fix(subos): 撤回 set 让位;只留 prepend 那半 —— 它才是 mcpp 的缺陷

上一版把 `set` 改成「用户已 export 就让位」。撤回,因为它错在两个层面:

**`set` 与「默认值」是两种意图。** 有些变量 subos 必须说了算 —— 指向它自己 loader
配置的那类,用户 shell 里一个陈旧值就能把环境弄坏。把两者塌成一个,等于从此无法
表达前者。

**op 词汇表是 xlings 的,不是 mcpp 的。** `envs` 是 xlings 的线格式;消费方悄悄给
一个 op 加第二种含义,会让同一个 subos 因为「由谁启动」而行为不同。

mcpp#382 想要的逃生通道应当是 xlings 新增一个 op(「声明默认值,用户可覆盖」),
mcpp 认它 —— 而不是在这里重解 `set`。现在未知 op 会被丢弃,所以那个 op 必须两侧
同时到位。已加断言把这个决定钉住,免得再被「修」回去。

留下的是 `prepend`:它字面意思就是接在已有值前面,而这里只发声明值、把调用方原有
内容整个丢掉。对 PATH 形状的变量,丢掉的是用户的整条搜索路径。这与 xlings 怎么修
无关,是 mcpp 自己的实现缺陷。

  tests/unit/test_subos_info.cpp  15 条(prepend 2 条 + set 语义 1 条)

Author: speak-agent

src/xlings/subos_info.cppm

@@ -256,18 +256,26 @@ resolve_env(const Info& info, const std::filesystem::path& subosDir,
             if (!hit) {
                 auto amb = ambient_of(d.var);
                 if (d.op == "set") {
-                    // A `set` is a DEFAULT, not an order. If the caller
-                    // already exported the variable, that wins.
+                    // `set` wins, ambient or not.
                     //
-                    // Recipes document this as the escape hatch -- xlings'
-                    // wsl-gl-host-link says in so many words that a user who
-                    // exports GALLIUM_DRIVER=llvmpipe keeps it. Before this,
-                    // the subos value overwrote it and the hatch did not
-                    // exist: `export GALLIUM_DRIVER=llvmpipe; mcpp run` still
-                    // ran with d3d12 and still failed (mcpp#382). An
-                    // environment a user set deliberately is the one piece of
-                    // input a build environment must not quietly overrule.
-                    if (amb && !amb->empty()) { out.emplace_back(d.var, *amb); continue; }
+                    // Deliberately NOT "yield to an exported value". That
+                    // reading was written here first and withdrawn: `set` and
+                    // "default" are two different intentions, and a subos has
+                    // real need of the first -- a variable naming its own
+                    // loader configuration must not be overridable by a stale
+                    // value in the caller's shell. Collapsing them here would
+                    // remove the ability to express it.
+                    //
+                    // It is also not mcpp's vocabulary to redefine. `envs` is
+                    // xlings' wire format; a consumer that quietly gives an op
+                    // a second meaning makes the same subos behave differently
+                    // depending on which tool launched the program.
+                    //
+                    // The escape hatch mcpp#382 asks for (a recipe declaring a
+                    // DEFAULT the user can override) therefore wants a new op
+                    // from xlings, not a reinterpretation of this one. When it
+                    // exists, it is honoured here -- unknown ops are dropped
+                    // today, which is why it has to arrive on both sides.
                     out.emplace_back(d.var, value);
                     continue;
                 }

tests/unit/test_subos_info.cpp

@@ -249,52 +249,8 @@ TEST(SubosInfo, UnknownOpIsDroppedLikeXlingsDrops) {
     EXPECT_EQ(env[0].second, "/yes");
 }
 
-// A `set` is a default; an exported value wins.
-//
-// The resolved pairs REPLACE the variable in the child (they go in as
-// extraEnv), so a `set` that ignores the caller's environment overwrites it
-// silently. Recipes are written on the opposite assumption: xlings'
-// wsl-gl-host-link documents that a user who exports GALLIUM_DRIVER=llvmpipe
-// keeps it, and that escape hatch did not exist -- `export
-// GALLIUM_DRIVER=llvmpipe; mcpp run` still ran with the subos value and still
-// failed (mcpp#382).
-TEST(SubosResolveEnv, SetDoesNotOverrideAnExportedValue) {
-    Tmp t;
-    t.write(R"({"workspace":{},"subos_info":{"schema_version":1,"runtime":"glibc@2.39",
-        "envs":{"glibc@2.39":[{"var":"GALLIUM_DRIVER","op":"set","value":"d3d12"}]}}})");
-    auto info = su::read(t.dir);
-    auto out = su::resolve_env(
-        info, t.dir, [](std::string_view v) -> std::optional<std::string> {
-            if (v == "GALLIUM_DRIVER") return std::string("llvmpipe");
-            return std::nullopt;
-        });
-    ASSERT_EQ(out.size(), 1u);
-    EXPECT_EQ(out[0].first, "GALLIUM_DRIVER");
-    EXPECT_EQ(out[0].second, "llvmpipe") << "the user's export must survive";
-}
 
-TEST(SubosResolveEnv, SetAppliesWhenNothingIsExported) {
-    Tmp t;
-    t.write(R"({"workspace":{},"subos_info":{"schema_version":1,"runtime":"glibc@2.39",
-        "envs":{"glibc@2.39":[{"var":"GALLIUM_DRIVER","op":"set","value":"d3d12"}]}}})");
-    auto info = su::read(t.dir);
-    auto out = su::resolve_env(
-        info, t.dir, [](std::string_view) { return std::optional<std::string>{}; });
-    ASSERT_EQ(out.size(), 1u);
-    EXPECT_EQ(out[0].second, "d3d12");
-}
 
-// An empty export is not an export.
-TEST(SubosResolveEnv, EmptyAmbientDoesNotBlockASet) {
-    Tmp t;
-    t.write(R"({"workspace":{},"subos_info":{"schema_version":1,"runtime":"glibc@2.39",
-        "envs":{"glibc@2.39":[{"var":"X","op":"set","value":"v"}]}}})");
-    auto info = su::read(t.dir);
-    auto out = su::resolve_env(
-        info, t.dir, [](std::string_view) { return std::optional<std::string>(""); });
-    ASSERT_EQ(out.size(), 1u);
-    EXPECT_EQ(out[0].second, "v");
-}
 
 // `prepend` prepends TO the caller's value rather than replacing it. Same
 // reason: the pair replaces the variable, so emitting the declared value alone
@@ -331,4 +287,27 @@ TEST(SubosResolveEnv, PrependIsIdempotentAgainstTheExportedValue) {
     EXPECT_EQ(out[0].second, existing);
 }
 
+// `set` wins over an exported value, and that is deliberate.
+//
+// This assertion exists to stop the opposite reading from being reintroduced
+// -- it was, once, as a fix for mcpp#382, and withdrawn: `set` and "a default
+// the user may override" are two intentions, and a subos needs the first for
+// variables naming its own configuration. The escape hatch that issue wants is
+// a NEW op from xlings, whose wire format this is, not a second meaning for
+// this one applied by one consumer.
+TEST(SubosResolveEnv, SetWinsOverAnExportedValue) {
+    Tmp t;
+    t.write(R"({"workspace":{},"subos_info":{"schema_version":1,
+        "runtime":"glibc@2.39",
+        "envs":{"glibc@2.39":[{"var":"GALLIUM_DRIVER","op":"set","value":"d3d12"}]}}})");
+    auto info = su::read(t.dir);
+    auto out = su::resolve_env(
+        info, t.dir, [](std::string_view v) -> std::optional<std::string> {
+            if (v == "GALLIUM_DRIVER") return std::string("llvmpipe");
+            return std::nullopt;
+        });
+    ASSERT_EQ(out.size(), 1u);
+    EXPECT_EQ(out[0].second, "d3d12");
+}
+
 }  // namespace